Human Motion Prediction via Adaptive Fusing Autoregressive and Non-Autoregressive Attention Networks

Human motion prediction is a challenging task that aims to predict future human poses based on an observed historical pose sequence. Previous approaches typically use an encoder-decoder model to extract spatiotemporal features of human motion, decode features and generate future poses either in an autoregressive or non-autoregressive manner. The former suffers from the issue of error accumulation, resulting in poor long-term predictions. The latter does not explicitly model temporal dependencies between predicted frames, which may adversely affect overall quality. To address the above issues, in this paper, we propose a novel hybrid framework combining autoregressive and non-autoregressive fashions. It employs a multi-branch parallel architecture to capture temporal dependencies at various time scales and the multi-head self-attention mechanism to encode spatial correlations among different body joints. Besides, we design a time-adaptive fusion module to integrate prediction outputs from each branch. We demonstrate the effectiveness of our proposed approach by surpassing previous state-of-the-art methods on two benchmark datasets. Code is available at https://github.com/aill/AFANAT.